orc-judge

Evaluates analysis, plan, and implementation artifacts at three ultra-lane judgment gates.

6|1|Updated Jul 4, 2026
One-click install
npx skills add https://github.com/azure-id/orc --skill orc-judge-azure-id
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: orc-judge
Source: https://github.com/azure-id/orc/tree/main/templates/skills/orc-judge
Command: npx skills add https://github.com/azure-id/orc --skill orc-judge-azure-id

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Deterministic checks like string comparisons and lint rules cannot judge whether an analysis interprets scope correctly, whether a plan is sound, or whether an implementation faithfully matches what the user asked for. This Skill adds a model-based quality gate on top of ORC's deterministic gates to catch those judgment-level failures. ## Core Features & Use Cases - Three dispatch contexts: gate=analysis (after the analyst), gate=plan (after the planner), and gate=implementation (after verify), each judging a different artifact against fixed ground truth. - Structured verdicts: returns APPROVE, REVISE, or ESCALATE with anchored, consequence-cited findings; REVISE loops the author with a hard 2-loop cap and a convergence rule. - Anti-anchoring design: the judge is read-only, blind to the author's self-assessment, and never fixes anything itself — authors fix. - Use Case: During an /orc-ultra run, after the planner finishes, the orchestrator dispatches the judge at gate=plan to validate decomposition, declared-file plausibility, and blast-radius coverage before any code is written. ## Quick Start Run /orc-ultra on a feature request and the orchestrator will automatically dispatch the judge agent at each of the three ultra gates.

Frequently Asked Questions about orc-judge

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I use the orc-judge skill?▼

You do not invoke it directly — it has no slash command. The judge is dispatched automatically by the orchestrator during /orc-ultra runs at three gates: after analysis, after planning, and after implementation verification.

What verdicts can the judgment gate return?▼

The judge returns APPROVE, REVISE, or ESCALATE. APPROVE continues the pipeline, REVISE sends blocking findings back to the author with a hard 2-loop cap, and ESCALATE presents the user with options like accept-and-proceed or stop-and-checkpoint.

Does the judge replace deterministic checks or user sign-offs?▼

No. Deterministic gates always run first, and the judge never re-checks what a string-compare catches. Judge approvals also never replace the analyst challenge round or the plan sign-off — they only add to them.

Can the judge fix the issues it finds?▼

No, the judge is strictly read-only and never fixes anything. Fixes are made by the original authors: the analyst for gate=analysis, the planner for gate=plan, and a scored executor fix wave for gate=implementation.

When is a blocking finding downgraded to advisory?▼

A blocking finding missing its verbatim anchor or class-appropriate justification is downgraded to advisory. The exception is a security finding with a concrete consequence, which is always blocking and never downgraded.